Question
What overall risk rating will Anthropic assign to the automated AI R&D threat model in the next Risk Report it publishes after 2026-08-28?
The dominant outcome, at 87%, is that Anthropic's next Risk Report will state the Automated R&D threshold has been met. This is driven by Anthropic's S-1/A No. 2 filing on November 10, 2026, which formally disclosed a company-level determination that the threshold has been met based on expected future model progress https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm. Since the Responsible Scaling Policy (RSP) explicitly defines the threshold as being met when Anthropic "observe[s] or expect[s]" the required doubling, this determination satisfies the policy's criteria. Reversing a formal securities disclosure in a subsequent RSP document would be legally and reputationally difficult. Comparison with broader trends in capability disclosures reinforces the primacy of this aggregate, expected-progress determination over individual-model capability disclosures in driving the threshold-met outcome.
The resolution criteria dictate that if the report states the threshold has been met, the question resolves to that specific option regardless of the overall direct catastrophic risk rating. This is a critical distinction because METR has noted that Anthropic's low/medium/high labels assess direct catastrophic risk, separate from the early-warning function of the threshold itself https://metr.org/blog/2026-11-05-rd-section-anthropic-risk-report-august-2026-review/. Therefore, even if Anthropic maintains a "Low" rating for direct risk, acknowledging the determination from the S-1/A will trigger the threshold-met resolution.
Substantive evidence also supports the threshold crossing, or at least the plausibility of Anthropic asserting it. Anthropic’s internal estimates of capability acceleration rose to 1.88× on November 9, with the system card promising further discussion on RSP implications https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card. METR’s independent assessment placed the speedup around 1.9× (with a 95% interval of 1.5–2.8×) and attributed roughly 80% to AI Independent Assessment of AI R&D Acceleration ... Independent Assessment of AI R&D Acceleration .... While no single released model crosses the threshold individually, Anthropic's stated reliance on company-wide R&D progress and expected future models provides the basis for their November 10 determination.
There are a few edge cases that account for the remaining probability. Anthropic has stated it is considering updates to the RSP Anthropic warns that it's very close to RSI capability, aka ... Anthropic S-1: AI Acceleration Threshold Met, Filing Says. A policy rewrite that redefines or splits the R&D threshold before the next report could lead to a conclusion that the newly defined threshold is not met, potentially prompting a "Moderate or higher" risk rating. Additionally, there is a minor risk that a report is published soon but uses a coverage date preceding November 10, technically avoiding the threshold-met statement for this cycle. I assign 5% to "Moderate or higher, but the RSP threshold is not met," 2% to "Low," and 1% to "Very low." These remain minor alternatives because neither a policy rewrite nor a coverage-date cutoff negates the weight of the aggregate determination already disclosed.
Finally, non-publication before the April 30, 2027 deadline presents a small but non-zero risk, to which I assign 5%. The RSP requires a Risk Report every 3–6 months Anthropic's Responsible Scaling Policy Anthropic: Responsible Scaling Policy — Version 3.4, placing the next due date around mid-February 2027. However, crossing the threshold triggers new external review requirements Responsible Scaling Policy. Bureaucratic friction from these reviews, internal evaluation redesigns, or IPO-related caution could potentially delay publication past the April cutoff, or result in an update that is not formally labeled a "Risk Report."
Comparing this with broader trends in capability disclosures confirmed that Anthropic's aggregate S-1/A expected-progress determination overwhelmingly drives the threshold-met resolution, leaving minor delay or policy rewrite edge cases largely unchanged.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited